A way of identifying a genuine series of images and a device for its implementation
The invention relates to devices and methods of identifying a genuine series of images. Its use in the operation of television equipment allows to obtain a technical result as facilitate the integration and timing of advertisements, videos, political speeches, etc., This result is achieved due to the fact that for individual images of the series of images to determine the characteristics of brightness, convert them into digital form and compared with the benchmark, while the signs of brightness decorrelator kvazistatichesky selection for many images. 2 C. and 30 C.p. f-crystals, 15 ill. The invention concerns a method of identifying a genuine series of images, such as advertisements, in which individual images of the series of images are selected, converted to digital form and compared with a reference sample attributes of their brightness.A series of images consisting of a set of forming a row and interconnected in the sense of individual frames should be identified, for example, in television at each new transfer. These series of images can be advertising messages, pay reports, old movies or video the x cases is the distribution of the series of images due payment, and legal or statistical reasons should be taken into account. The authenticity of the series of images means that the content of these images are saved in its original form, i.e. all the pixels of the image according to its brightness and color are unique and therefore unchanged. When you copy a technically high-quality equipment this unique location can not be changed and supplemented according to any parameter. The notion of "authentic" refers, therefore, to the copied series of images.Television stations advertising messages are transmitted, for example, at this time, which is of particular interest to advertisers. For advertisers, it is important to ascertain whether or not his message is transferred at the appointed time. During its existence, the advertising message changes its appearance, i.e., it is shortened, some frames are modified or completely replaced. This new series of images to be distinguished from the original version.Proposal from Germany 4309957 C1-known way of identifying a genuine series of images. While the individual elements of the image or the brightness of which is converted into digital form and quality characteristics compared to the characteristics of known frames of reference samples.In connection with the necessity of compression or compaction data, the probability of misidentification is relatively high. Therefore, to reduce the number of erroneous identifications in the known methods are scanning multiple, sequentially following frames. However, the drawback is that successive frames are usually great similarity, which signs often lead to accidental similarity without really similar images. These so-called accidental misidentification, however, does not necessarily negatively affect the identification of the advertising message as such, but lead to a large amount of data that load computing device.The present invention is therefore the reduction of the amount of data and reducing the number of erroneous identifications.According to the invention this problem is solved due to the fact that the signs decorrelate their kvazistatichesky selection through multiple frames (images).The result is that, instead of a selection of signs adjacent frame signs are selected kvazistatichesky way, bypassing several frames, it becomes possible to neglect the case is the result of the reduced number of erroneous identifications and associated loads computing device, the amount of data.According to a preferred variant of the invention, the selected characteristics of the individual frames are recorded in an ordered sequence in the storage device of the carousel images, made in the form of a shift register, and read kvazistatichesky way with a custom handle. Turning to characteristics occurs when the maximum pass images. At the same time reach out to the signs with a maximum interval between frames. Thanks to such extent provided decorrelation characteristics as dissolved temporal and spatial relationships.According to another preferred variant of the invention is a two-step processing method, in which the first stage approximate and second accurate identification. In terms of data preliminary identification reduced so that it is possible in real-time. It allows you to recognize the advertising message without identifying possible distortions. Accurate identification is carried out only in the case when the approximate identification of a series of images or advertising message will be correlated with existing atalonia and temporal distortions.According to another preferred variant of the method for the formation of signs used to change the brightness within a cluster formed by spatially related pixels in the image.In this case, the clusters are subjected to discrete cosine transformation. One of the coefficients of low-frequency alternating reduced to his mark and is used as the characteristic. A series of images is divided into time slots of equal duration, where each quantum is an independent, correlated unit. Thus, the frames of one time slot encoded at extreme seal data. Identification or, more accurately approximate identification of the advertising message occurs only when the individual intervals are identified by their proper sequence and the proper time interval.The invention concerns also a device for implementing this method.Proposal from Germany 4309957 C1 is known a device for implementing the method of identifying a genuine series of images. This device has the disadvantage that there is only one selection of attributes consistently sent a large number of erroneous identifications.Therefore, another objective of the present invention is a device that provides an opportunity kvazistaticheskogo selection of signs.This problem is solved due to the fact that the input of the shift register, made in the form of a storage device, through the transducer DCT (discrete cosine transform) is connected to the decoder that contains the raster Converter/cluster, and the output is connected through the correlator to the reference memory with the possibility of presenting a series of images coming from the receiver to the video decoder, the correlator in the form of a characteristic vector and the comparison with the reference pattern stored in the reference memory.Thanks to the use of the shift register becomes possible in a simple way to produce quasistatically selection of attributes, while decorrelate signs of different frames (images).According to another preferred embodiment of the invention between the DCT Converter and a shift register is provided by the branch in the direction to a storage device data accurate identification, which is made in the form of a shift register FIFO (first in, first out) (first input first output) and contains all the features that characterize carisle conducted in real-time approximate identification.Other details of the invention are described below in the detailed description and the attached drawings in which is illustrated as an example of preferred embodiments of the invention.In the drawings shown: Fig.1 is a block diagram of an identification device of a genuine series of images, Fig. 2 scheme of advertising messages in the form of a series of images with decomposed into segments of time series of images, Fig.scheme 3 length of time with symptoms (brightness), Fig.4 the feature vector length of time in Fig.3 in the form of a binary stream, Fig. 5 is an example illustrating the distortion in which the product has lost the definition of "new", Fig. 6 is an example illustrating the local distortion within a single frame, which changed the writing of the letter (fragment), Fig. 7 spectral reference samples in the form of the base image with the least significant coefficients, Fig. 8 options of the individual coefficients of the discrete cosine transform, Fig.9, the seal data of the brightness signal in CCIR format,
Fig. 10 the density of the amplitude distribution of the coefficient of any frame without Downsampling (abbreviated sample),
Fig. 11 the density of the amplitude distribution of the coefficient C01Uravnenii with theoretically expected binomial distribution
Fig.13 scheme carousel images,
Fig.14 the distribution of random matches binary samples at different dekoriruemyh actions,
Fig. 15 the effect of dekoriruemyh action when the threshold is exceeded, the identification of the fragment of Fig.14.Device to identify a genuine series of images 1 consists of a video decoder 2 Converter for raster/cluster, whose input 3 is connected to the output of the receiver 4. The output of decoder 2 is connected to the input of the Converter D 5. The output of the DCT Converter 5 is connected to the input 6 of the carousel 7 images. Register 7 offset images is associated with its exit 8 to the input of the correlator 9. The correlator 9 is connected with the reference memory 10. The correlator 9 is connected with the analyzer 11 approximate identification. Between DCT Converter 5 and case 7 shift is a shift register FIFO 12 associated with the analyzer 13 for accurate identification. The analyzer 11 approximate identification is the first stage of the signal processing approximate identification. If the analyzer 11 is detected a series of 14 images, which has great similarity with stored in a frame memory 10 reference sample, and if with high probability (for example, probability &g is eficacia receives the signal from the analyzer 11 approximate identification and on the second stage of processing is in effect, using a stored in the shift register FIFO 12 data. Accurate identification ensures ultimate confidence and finds a local and temporal distortions.Due to temporary distortions (cuts or scene changes), you must encode the entire series of 14 images. For this purpose it is expedient to divide a series of 14 images at time intervals of 15 duration, for example, 2 seconds. Time segments 15 are independent units, each of which is correlated separately. Identification (approximate) of a series of 14 images occurs when individual time segments 15 are identified in the correct order and the correct time interval with respect to the analyzer 11, 13. When there is a temporary distortion will be no separate time periods of 15, while the rest will be identified in the expected sequence. In numerous experiments mentioned duration of about 2 seconds was determined as optimal, but it can be adjusted to ensure that advertising messages or series of 14 images, set the preferred duration, and then periods of time 15 pick with this asininely duration of the advertising message 7 seconds then the length of the individual time slot 15 is equal to 1.75 seconds. Otherwise, when the correlation will remain excluding indivisible then the rest.Local distortions, i.e., changes within a frame 16 of a series of 14 images, cover area, expressed in terms of image. In Fig. 6 shows an example in which changed the writing of the letter (fragment). Significant changes take area equal to 32x32 point. When accurate identification, i.e. identification of likely available local distortions, each frame is divided into sections of approximately the specified values to be correlated (compare) in the processing device 13 as independent units with corresponding parts of other frames. To encode one group of image points, i.e., one cluster variable is the share of luminosity or brightness within the same cluster. This variable percentage points - completely in the General plan to change the brightness within the cluster, as opposed to a constant fraction of which brightness is average. The brightness change can occur in different ways: it can act as a simple gradient with a specific direction or the performed by the DCT Converter 5, was used to detect the alternating bright and dark. Such discrete cosine transform is applied also in a known density algorithms, such as JPEG, MPEG, and so embedded in highly integrated chips, such as Zoran 36050. When such sealing frame, the frame is divided into clusters (blocks) of size 8x8 pixels (Raster to Clustar-Conversion(converting raster/cluster), which are then subjected dvuhratnomu discrete cosine transform. In this way we obtain a frequency view images.After that, the seal principle is that high frequency in the share of heavily compacted or even lower. Spectral reference samples for individual coefficients Cnkcalled basic frames, the minimum significant of which is shown in Fig.7. Top left base frame (factor C00- when using cluster 8x8 points - provides a constant share, i.e., the average brightness of the cluster. Top right base frame (factor C01) is used to check the extent to which shows the characteristic brightness is present in the studied cluster. Therefore, it provides the lower the provide relevant information in the other direction of the frame. Base frames for DCT coefficients of higher order are not given, since the present invention they are not important. Discrete cosine transform is characterized by the property that at which substantial information, distributed in the area between the original all auxiliary values after conversion is concentrated in a few components, i.e., significant energy shares on the so-called DC coefficients (constant fraction) and on the lower coefficients AC (share of low-frequency alternating).In Fig. 8 shows the decrease in the energy in the AC coefficients of higher order. The share of energy, for example, the coefficient C01it is advisable to approximately doubled if the horizontal and vertical is the decimal reduction. Under this you should understand averaged four, square-spaced image points in one new point. This method is also called Downsampling is downsampled. In this way, the cluster size of 16x16 pixels is reduced to a size of 8x8 pixels, to which you then apply a discrete cosine transform. Received now the coefficients Coicharacterized, as a rule, udomi means this method is partly common in television by skipping rows, when images are sent by fields, lines and in the form of a comb. If you consider only the field that will be present vertical downsampled. Horizontal downsampled can be further obtained using chips commercially available. In Fig.10 shows the density of the amplitude distribution of the coefficient Coifor any frame 16 without Downsampling. In Fig.11 shows the density of the amplitude distribution of the coefficient of Coi in Fig.10 with Downsampling. The wider the curve in Fig. 11 indicates in figures almost double the amount of deviation from the mean compared to the curve in Fig.10.After all the frames 16 time 15 decomposed into clusters of size 8x8 pixels after the decimal reduction they are blocks of images or clusters of size 16x16 pixels and converted into spectral region, the lower the coefficient of active rotation - it is C01or10each cluster is further increasing due to the fact that in the future, use only its sign. Therefore, the cluster size 8x8 pixels present in the form of a bit indicating the sign of the coefficient of alternation. Bliod the entire series of images is encoded at extreme seal data. Each bit is a local characteristic that does not depend on the modulation of the image signal and the ratio signal/noise.Seal data of the brightness signal format CCIR average of 768x576 pixels, reaches first about 2103. The duration of the period of time equal to, for example, 2 seconds, the amount of data is still about 11 KB. When processing in real time thousands of advertising messages that amount of data would be too large. Therefore, the approximate identification applied as suitable for real-time identification of a series of images without it occurred distortion. Accurate identification is then superimposed on the approximate, being unsuitable for real-time way, confirms the results and allows to analyze the distortion. Approximate identification is already possible in the presence of about 16 characteristics 17, 18 brightness for each frame 16. Under the sign 17, 18 brightness, you should understand the sign bit of the DCT coefficient. Thus according to a certain scheme from 1728 significant bits of one (semi) frame selected respectively 16 bits, i.e. two bytes. At two-second duration of the segment time is impressive. The video objects are, as a rule, the length, including many clusters. For these clusters are likely the same DCT coefficients. Signs 17, 18 brightness selected from these clusters, therefore, are not independent from each other and do not improve the quality of identification. Relationships are destroyed when through kvazistaticheskogo method selected from the clusters, as remote from each other, 16 signs 17, 18 brightness. In this case, do likewise in respect of each of the frame 16, and generally get the data chain length, for example, 1650=800 bits. Assuming independence of the individual bits, then the probability of a random match of two such binary combinations in to bits of N, it is possible according to the binomial distribution:
where p is the probability of occurrence of the characteristic 17, 18 brightness.In the cases R=0,5, i.e., 0 and 1, being significant bits of the coefficient, likely to the same extent. Then the binomial distribution will be reduced to
b(k, N, p) = (N)pN.If you set the lower limit of identification, equal to 85%,
all of the similar elements is tion is related to the true identification of the series of images is determined by monitoring the reliability, incorporated in the software. Otherwise, it may be a random match, the so-called erroneous identification.To prevent erroneous identifications provided by the presence of the register 7 offset. Register 7 offset is a storage device 19 in the form of a shift register, which records selected characteristics 17, 18 of the brightness of the frame 16 in an ordered sequence. The purpose register 7 offset images is such processing multiple frames 16, in which the samples are not based on a sequence of frames, and quasilocality the sample through multiple frames 16. Is the so-called interframe coding. In Fig.14 depicts the distribution of the random matching such interframe coding, and processing was carried out after 2, 10 and 50 images. Curve 1 shows the so-called interframe coding, in which the signs are processed in the sequence of frames. In Fig.15 in the enlarged view shows the area above the threshold identification equal to 85%. After 40 MS signs 17, 18 brightness of a new image 20 is entered in the storage device 19, while the "oldest" image 21 is displayed naruhodo kvazistatichesky way in selective conversion of N-signs 17, 18 brightness period of time and served in the correlator 9 for comparison with a reference sample. When this act so as to avoid secondary reading cluster of the frame 16 when it is placed in register 7 offset images. Turning to characteristics 17, 18 brightness is at maximum skipping frames 16 and at the same time subject to the maximum interval between frames 16. Due to this effect is achieved by the decorrelation characteristics 17, 18 brightness as dissolved temporal and spatial relationships. The method can be performed by the following operations:
a) a Series of 14 images are divided into time segments 15 duration 1.5-2 seconds (according to task identification).b) Each frame 16, according to the JPEG method, is divided into clusters, and it is preferable to apply vertical and horizontal downsampled. Clusters of frames subjected to discrete cosine transform, and apply the sign of the minimum significant ratio of alternation as element 17, 18 brightness (preferably C01or C10).From 1728 signs 17 created according to p. b) for each frame 16, the selected subset of about 800-1000 signs 18 and correlate the tool decorrelates method of treatment selected about 800-1000 signs 18 1 period of time of 15, and lay in the form of a vector of 22 signs, inherent throughout the period of time. This corresponds to approximately 16-32 bits per frame.d) the feature Vector 22, obtained in paragraph (C), denoted as the reference sample and subsequently, when the work is compared with the test samples obtained continuously in a similar way on the basis of the transmitted program, the reciprocal link EXNOR. When the similarity is greater than a specified threshold (e.g., 85%), identification is made. Identified objects, their degree of similarity, are brought together with the exact time stamp of the frame in a data Bank containing the appropriate zone compared with all reference samples.d) data Bank constantly monitored by the relevant software on the relationship of the time segment 15, which is identified at the proper interval between shots and when there is sufficient similarity. In the case of the identification of the majority of the time segment 25 of a series of 14 images the entire series of images is identified.e) the Process of identification pursuant to sub. g), d) occurs in real time, resulting in the identification of a series of images can be shown directly after the transfer of the series is 2 - this 86400 bits=10.8 KB-that is temporarily stored in the shift register FIFO 12, serves for accurate identification. Here is the correlation of all divided by the area frame 16 all time intervals of 15 series of 14 images. This can be compared with the pre-selected reference samples, as the identification of a series of images actually happened.Captions to figures
Fig.2: 1 - advertising messages, 2 losst<2 c.Fig.3: 1 - image.Fig.10: 1 - without a short sample (Downsampling), 2 - frequency of occurrence, %, 3 - indicators DCT, 4 - deviation from the mean.Fig.11: 1 - reduced sample, 2h2v, 2 - frequency of occurrence, %, 3 - indicators DCT, 4 - deviation from the average value, 5 - factor C01.Fig.12: 1 - binomial distribution, 2 - a carousel of images, 3 - without the carousel images.Fig.13: 1 - read.Fig.14: 1 - internal frame, 2 - 2 interframe image, 3 - 10 interframe image, 4 - 50 interframe image.Fig.15: 1 - internal frame, 2 - 2 interframe image, 3 - 10 interframe image, 4 - 50 interframe images, 5 - threshold identification, 80%.
2. The method according to p. 1, wherein the selected characteristics (17, 18) the brightness of the individual images recorded in an ordered sequence in a memory device (19) of the casing (7) of the shift and read kvazistatichesky way with a custom handle.3. The method according to p. 2, wherein accessing characteristics (17, 18) brightness is at the maximum possible pass (23) images (16).4. The method according to p. 2 or 3, characterized in that the reference to signs (17,18) the brightness is at the maximum possible interval within images (16).5. The method according to one of paragraphs. 2-4, characterized in that after about 40 MS symptoms (17, 18) brightness new image (20) is introduced into the storage device (19), and signs (17, 18) the brightness of the earliest images (21) make it to the end of the storage device (19).6. The method according to one of paragraphs. 2-5, characterized in that the signs (17, 18) brightness served in the correlator (9) for comparison with the benchmark.7. The method according to one of paragraphs. 2-6, characterized in that the series (14) images Razak time (15) is an independent, correlated unit.9. The method according to p. 7 or 8, characterized in that the series (14) of the images is identified, if identified separate periods of time (15) in the sequence corresponding to the reference images and in the time interval corresponding to the reference images.10. The method according to one of paragraphs. 7-9, characterized in that the series (14) images are divided into periods of time (15) duration 1.5-2 C.11. The method according to one of paragraphs. 7-10, characterized in that the duration of the time segment (15) is chosen such that the duration of the series (14) images could be divided without a remainder for the duration of the segment (15).12. The method according to one of paragraphs. 1-11, characterized in that the image (16) is divided into many clusters formed spatially related pixels in the image.13. The method according to p. 12, characterized in that for the formation of symptoms (17, 18) brightness use the brightness change of the cluster.14. The method according to p. 12 or 13, characterized in that the clusters are subjected to discrete cosine transform.15. The method according to p. 14, characterized in that the quality characteristics (17, 18) brightness use the coefficients of the discrete cosine transform.
The formula soborom for individual images of the series of images to determine the characteristics of brightness, convert them into digital form and compared with the benchmark, characterized in that the signs (17, 18) brightness decorrelator kvazistatichesky selection for many images (16).
17. The method according to p. 16, characterized in that the ratio of low-frequency alternating limit its mark and is used as a characteristic(17, 18) brightness.18. The method according to one of paragraphs. 12-17, characterized in that the cluster size is 88 points in the image.19. The method according to one of paragraphs. 12-18, characterized in that the clusters are subjected to subdirectly by the decadal decrease in their number horizontally and vertically.20. The method according to p. 19, characterized in that four arranged in a square image pixels are combined into one new point in the image.21. The method according to p. 20, characterized in that the cluster size 1616 points reduce to size 88 pixels.22. The method according to one of paragraphs. 1-21, characterized in that conduct a two-stage processing, which at the first stage, the approximate identification, and the second stage is an exact identification.23. The method according to p. 22, characterized in that the amount of data during the approximate identification of the cut so that it becomes possible in real time.24. The method according to p. p. 22 or 23, characterized in that for the approximate ID is moneysee device (19) of the casing (7) shift to correlate in real time, select decorrelation method of treatment process and then in the form of vectors (22) signs each time interval (15).25. The method according to p. 24, characterized in that the vector (22) signs of the time segment (15) is compared with the corresponding reference link is EXCLUSIVE-NOT-OR and in identifying contribute in a data Bank containing the appropriate fields for all of the compared reference samples.26. The method according to p. 25, characterized in that the data Bank is constantly in control of appropriate software for the presence of interconnected lengths of time (15), which were identified at the correct interval between images with sufficient affinity.27. The method according to p. 26, characterized in that the series (14) of the images is identified, if identified most time periods (15) this series (14).28. The method according to one of paragraphs. 22-27, characterized in that the exact identification is performed only in the case when the series (14) images after approximate identification corresponds to the available reference sample.29. The method according to p. 28, characterized in that when the accurate identification of the process all the signs (17, 18) brightness, forming parts of the image is a (7) shift, made in the form of a storage device (19), through the DCT Converter (5) is connected to the decoder (2), containing the raster Converter/cluster, and the output is connected through the correlator (9) to the reference memory (10) with the possibility of filing series (14) of the images coming from the receiver (4) to the video decoder (2), the correlator (9) in the form of a vector (22) characteristics and comparison with standard stored in the reference memory.31. The device according to p. 30, characterized in that between the DCT Converter (5) and case (7) shift provided by the branch in the direction to a storage device for accurate identification.32. The device according to p. 31, wherein the storage device accurate identification is made in the form of a shift register FIFO, which stores all the signs (17) brightness series (14) images.
FIELD: automated recognition of symbols.
SUBSTANCE: method includes following stages: tuning, forming symbols models, recognition, recording background model together with background of read image, separating model of registered background from elementary image of background, combining for each position of symbol of model of letters and/or digits with elementary displaying of appropriate background, forming of combined models, comparison of unknown symbols to combined models, recognition of each unknown symbol as appropriate symbol, combined model of which is combined with it best in accordance to "template comparison" technology.
EFFECT: higher efficiency.
10 cl, 10 dwg